Image Processing Reference
In-Depth Information
The fused image is generated from a combination of these spatially similar weights
which preserves the discontinuities in the data. The fused image, thus, possesses
very high values of contrast and sharpness without driving the individual pixels
into over- or under-saturation. By using an appropriate auxiliary variable, we show
how the constrained optimization problem can be converted to a computationally
efficient unconstrained one.
We have already discussed the band selection scheme based on the additional
information or the conditional entropy of the successive hyperspectral bands. This
scheme exploits the redundancy in the hyperspectral bands in order to select a
specific subset of mutually less correlated bands. Consider the optimization-based
fusion technique which generates a fused image that possesses certain desired
characteristics. While we have explored an output-driven fusion method, one may
as well want to investigate a band selection strategy based on the output fused
image. We explore the idea of the selection of the specific hyperspectral bands
based on the fusion output. This band selection scheme selects a band for fusion
if the given band possesses significant additional information as compared to the
fused image obtained by combining all bands selected so far. In other words, a given
band is selected for fusion when the conditional entropy of the band with respect to
the corresponding intermediate fused image exceeds a certain threshold. Naturally,
this band selection scheme is specific to the fusion technique to be employed,
which also governs its performance. In this scheme, the conditional entropy of
the band is evaluated against the output image, rather than input bands. Therefore,
if the band is visually quite different from the present output, it get included in
the subset that produces the subsequent fused image. We can, therefore, obtain
the fused image which captures most of the independent information across the
bands using a quite less number of bands. The objective assessment of the fused
image, however, shows a very minimal degradation in the quality of the image as
compared to the image obtained from fusion of entire hyperspectral data using the
same fusion technique.
While the techniques for fusion of hyperspectral images are being developed, it is
also important to establish a framework for an objective assessment of such fusion
techniques. Such an assessment provides uniformity in the process of evaluation
across various fusion techniques, and is also useful to explore various salient
characteristics of the fusion technique. The extension of existing measures towards
the evaluation of fusion of hyperspectral images is a non-trivial task. This problem
is difficult due to a very large number of input bands, and absence of any reference
image or the ground truth. We explain how one can extend some of the existing
measures for evaluation of hyperspectral image fusion techniques. We also discuss
several modifications in some of the measures for a better and efficient analysis
of the fusion process. As a large number of techniques for generalized image
fusion have been developed, one would like to extend these for the fusion of a
very large number of images such as the hyperspectral data. We explain the notion
of consistency of a fusion technique as more and more images are being fused
using the same fusion technique. This consistency analysis can be very useful in
Search WWH ::

Custom Search